🎯 Goal

In this project, we aim to build a modular deep learning system for multi-task learning that can simultaneously perform disease grading (multi-class classification) and lesion segmentation (binary or multi-label) from retinal images. The system should be capable of handling both single-task learning and multi-task learning scenarios.

📁 Dataset

We utilize the Indian Diabetic Retinopathy Image Dataset (IDRiD), which can be downloaded here. Make sure the dataset follows this structure.

dataset/
├── Disease_Grading/
│   ├── ground_truth/
│   │   ├── disease_grading_train.csv
│   │   └── disease_grading_test.csv
│   └── original_images/
│       ├── Train/
│       └── Test/
├── Segmentation/
│   ├── ground_truth/
│   │   ├── Train/
│   │   │   ├── Haemorrhages/
│   │   │   ├── Hard_Exudates/
│   │   │   ├── Microaneurysms/
│   │   │   ├── Optic_Disc/
│   │   │   └── Soft_Exudates/
│   │   └── Test/
│   │       ├── Haemorrhages/
│   │       ├── Hard_Exudates/
│   │       ├── Microaneurysms/
│   │       ├── Optic_Disc/
│   │       └── Soft_Exudates/
│   └── original_images/
│       ├── Train/
│       └── Test/

🚀 Train and Validate

To train and evaluate the model, run the following command with the appropriate config path:

bash experiment.sh

📊 Results

Model	Classification Loss	Segmentation Loss	Accuracy (%)	Dice Score
Classification Only	1.4005	0.0000	60.19	NA
Segmentation Only	NA	0.0261	NA	0.6806
Multi-task	1.5739	0.0178	62.14	0.7029